This study examines socially-conditioned variation within semantic domains like kinship and weather using thirteen Indian cities as a case-study. Using bilingual social media data, we infer six semantic domains from corpora representing individual cities with a lexicon including terms from English, Hindi and Transliterated Hindi. The process of inferring semantic domains uses character-based embeddings to retrieve nearest neighbors and Jaccard similarity to operationalize the edge weights between lexical items within each domain. These representations reveal distinct regional variation across all six domains. We then examine the relationship between variation in semantic domains and external social factors such as literacy rates and local demographics. The results show that semantic domains exhibit systematic influences from sociolinguistic factors, a finding that has significant implications for the idea that semantic domains can be studied as abstractions distinct from specific speech communities.